THE HUMANPEN BLOG

Clear answers on AI detection and academic writing

Research-backed guides on how detectors work, how to read your Turnitin report, and how to keep long documents intact.

AI detectorsFeatured

Why non-native English writers get flagged by AI detectors more often

Seven detectors called 61% of human-written TOEFL essays AI. The reason is not cheating - it is that detectors measure vocabulary range, and a controlled experiment in the same paper proves it.

· 9 min read

61.2%

of human-written TOEFL essays misclassified as AI, averaged across seven detectors

AI detectors

What AI detectors actually measure: perplexity and burstiness

Two statistics, neither of which is authorship. And the most-cited detector stopped using both of them in 2023 - which changes what you can infer from a score.

· 7 min read

AI detectors

Why the same text scores differently on every detector

In the Stanford study, 89 of 91 essays were flagged by at least one detector and only 18 by all seven. Same essays, same day - the tools disagreed about 71 of them.

· 6 min read

Turnitin reports

How to read a Turnitin AI writing report

The percentage is not the share of your document that is AI. It is the share of one specific subset of it - and knowing which subset explains most of what looks wrong in these reports.

· 8 min read

Turnitin reports

Is 20% AI too high? Why there is no absolute threshold

No number is universally safe, and the vendor will not publish one below 20 percent at all - because that is the range where its own testing found the most false positives.

· 7 min read

Turnitin reports

Flagged, but you wrote it yourself: preparing a response

A detector score is not proof, and the bodies that review these cases have said so. What decides an appeal is usually provenance and procedure, not the number.

· 8 min read

Academic writing

Choosing a citation style by discipline: a comparison table

Which style your field expects, what actually distinguishes them, and the two that trip people up - Harvard, which has no single standard, and Vancouver and AMA, which look identical and are not.

· 8 min read

Academic writing

Publisher AI disclosure policies, compared

Every major publisher permits AI assistance and requires you to declare it - but each wants the declaration in a different place, and each exempts something different. Checked 2 August 2026.

· 7 min read

Tools & workflow

Why Zotero and EndNote exports never quite match the required style

The style file is usually fine. What is missing is the metadata it needs - and one field in particular is, in Zotero’s own words, the most inconsistently provided field across publishers.

· 7 min read

Tools & workflow

Word fields, tables of contents and cross-references: why copy-paste breaks a thesis

A table of contents is not text. Neither is "as shown in Figure 3". They are fields - and a plain-text round trip destroys every one of them silently.

· 7 min read